From what I read on the SETI boards Tesla is not yet supported even in SETI. I see a lot of confusion going on CUDA. Also the ATI people would like to use their graphic boards. I think the GPUs are still in beta phase.
Tullio
At the moment CUDA can only be used on two projects SaH and GPU Grid ... the SaH implementation has some issues and the general feeling by the mega-crunchers there is that the results returned by the CUDA application have a higher rate of error ...
GPU Grid on the other hand the application is, or appears to be more stable though the current version has a problem with priority and there are server issues on getting work ...
Add in the problems with the work fetch and resource share and you have many potential problems.
If, however, you are not adverse to baby sitting your computer some then going with GPU Grid is not a bad choice at the moment.
On a 9800 GT card the tasks take a minimum of 6 hours with 17 being nominal (rough numbers, it has been a couple days) ... with a GTX 280 it is about 4-6 hours ...
If you are buying a new GPU to process the cost benefit is for the GTX 260 card with the additional processing elements. GPU Grid has some involved discussions ...
In any case, if you want to be an early adopter you can start there and when Einstein gets off the dime you will be set ... :)
Sorry guys, I could not resist ...
Seriously, it is very early days and if you are not up to the hassles of early adoption, then there is no reason not to wait a bit for the dust to settle. THough version 6.5.0 is working reasonably well for me there are serious issues with the BOINC Manager and the way that it does not properly manage resources ...
At the moment CUDA can only be used on two projects SaH and GPU Grid ... the SaH implementation has some issues and the general feeling by the mega-crunchers there is that the results returned by the CUDA application have a higher rate of error ...
I believe there is also Folding@Home, although i've never tried it myself.
Bit of an understatement there about "higher rate of error".
Quote:
GPU Grid on the other hand the application is, or appears to be more stable though the current version has a problem with priority and there are server issues on getting work ...
Add in the problems with the work fetch and resource share and you have many potential problems.
If, however, you are not adverse to baby sitting your computer some then going with GPU Grid is not a bad choice at the moment.
On a 9800 GT card the tasks take a minimum of 6 hours with 17 being nominal (rough numbers, it has been a couple days) ... with a GTX 280 it is about 4-6 hours ...
If you are buying a new GPU to process the cost benefit is for the GTX 260 card with the additional processing elements. GPU Grid has some involved discussions ...
In any case, if you want to be an early adopter you can start there and when Einstein gets off the dime you will be set ... :)
Sorry guys, I could not resist ...
Seriously, it is very early days and if you are not up to the hassles of early adoption, then there is no reason not to wait a bit for the dust to settle. THough version 6.5.0 is working reasonably well for me there are serious issues with the BOINC Manager and the way that it does not properly manage resources ...
Yes the joys of a "development" version. Still it does handle CUDA quite nicely, now if they could fix up the other bits they broke in the process...
Yes the joys of a "development" version. Still it does handle CUDA quite nicely, now if they could fix up the other bits they broke in the process...
NOt sure about the bits they broke in the process, but it does not handle CUDA well ... it lumps the GPU in with the CPUs and you can get a situation where the GPU is dry and BOINC Manager will ask for 0 seconds because the Resource Share and work fetch do not track the GPU tasks separately from the CPU tasks ...
It remains to be seen if any of the other issues raised are addressed at all in the next couple of versions. Based on the trend of the discussion it does not look to be so ...
Anyway, with "fiddling" I have been able to keep my GPUs busy and just went over 100K CS for GPU Grid this evening ... onward to 400,000 ... maybe as soon as the end of the month? Sooner now I have both GPUs working again ...
Yes the joys of a "development" version. Still it does handle CUDA quite nicely, now if they could fix up the other bits they broke in the process...
NOt sure about the bits they broke in the process, but it does not handle CUDA well ... it lumps the GPU in with the CPUs and you can get a situation where the GPU is dry and BOINC Manager will ask for 0 seconds because the Resource Share and work fetch do not track the GPU tasks separately from the CPU tasks ...
It remains to be seen if any of the other issues raised are addressed at all in the next couple of versions. Based on the trend of the discussion it does not look to be so ...
Well 6.5.0 working within the bounds of the current design is better at cuda than 6.4.5. The bits they broke, well don't try to terminate the science apps via BOINC.
I put my 2 cents worth in regarding resource share before attempting to address work-fetch policy. No doubt they will be ignored or put in the "too hard" basket.
I put my 2 cents worth in regarding resource share before attempting to address work-fetch policy. No doubt they will be ignored or put in the "too hard" basket.
I think they are being put into the category of: "I am not going to do it because I really don't care" ...
But that is just my bitter, cynical, self speaking ...
None of the issues in BOINC are really all that hard ... there is just a lack of interest in NIH ideas ... so, when we peel back the curtain of time we find all kinds of fixes some of them even coded and tested (I know, I tested them) that never made it into the code base (as far as I know, they may have been snuck in while I was not doing BOINC, though I suspect not because I am still seeing "funky" behavior in task scheduling) ... because Dr. Anderson did not want them.
It mattered not that it addressed problems that were particular with multi-core systems (less evident on dual cores, noticeable on 4 core and better) ... but I digress ...
I put my 2 cents worth in regarding resource share before attempting to address work-fetch policy. No doubt they will be ignored or put in the "too hard" basket.
I think they are being put into the category of: "I am not going to do it because I really don't care" ...
But that is just my bitter, cynical, self speaking ...
None of the issues in BOINC are really all that hard ... there is just a lack of interest in NIH ideas ... so, when we peel back the curtain of time we find all kinds of fixes some of them even coded and tested (I know, I tested them) that never made it into the code base (as far as I know, they may have been snuck in while I was not doing BOINC, though I suspect not because I am still seeing "funky" behavior in task scheduling) ... because Dr. Anderson did not want them.
It mattered not that it addressed problems that were particular with multi-core systems (less evident on dual cores, noticeable on 4 core and better) ... but I digress ...
When I was at Seti we physically talked to Dr. A and he was not interested in whatever we had to say. He has a path he is on and will not be deviated from it. IMO it will lead to a different program, other than Boinc, in the future if he does not let change happen. I am not saying this will happen net week, but people will continue to leave if things don't get better at some point.
On a more positive note I bought a 9600 GSO card and it is taking well over 13 hours PER UNIT on Gpugrid. And that is only a projection, it has not actually finished even one unit yet and has run 24 hours! It is in a dual core, 2.66ghz, Vista Home Basic system. I downloaded the latest drivers from Nvidia, but currently it is NOT better!
I put my 2 cents worth in regarding resource share before attempting to address work-fetch policy. No doubt they will be ignored or put in the "too hard" basket.
I think they are being put into the category of: "I am not going to do it because I really don't care" ...
But that is just my bitter, cynical, self speaking ...
None of the issues in BOINC are really all that hard ... there is just a lack of interest in NIH ideas ... so, when we peel back the curtain of time we find all kinds of fixes some of them even coded and tested (I know, I tested them) that never made it into the code base (as far as I know, they may have been snuck in while I was not doing BOINC, though I suspect not because I am still seeing "funky" behavior in task scheduling) ... because Dr. Anderson did not want them.
It mattered not that it addressed problems that were particular with multi-core systems (less evident on dual cores, noticeable on 4 core and better) ... but I digress ...
When I was at Seti we physically talked to Dr. A and he was not interested in whatever we had to say. He has a path he is on and will not be deviated from it. IMO it will lead to a different program, other than Boinc, in the future if he does not let change happen. I am not saying this will happen net week, but people will continue to leave if things don't get better at some point.
On a more positive note I bought a 9600 GSO card and it is taking well over 13 hours PER UNIT on Gpugrid. And that is only a projection, it has not actually finished even one unit yet and has run 24 hours! It is in a dual core, 2.66ghz, Vista Home Basic system. I downloaded the latest drivers from Nvidia, but currently it is NOT better!
I was pointed to a site by a comment elsewhere and it indicated a BOINC Manager branch that was now being independently maintained. I don't know how good it is or even if it is active at this time ... but I did find that encouraging. I have been coming to the feeling that the only way things are going to look up for BOINC is for it to lose all funding ... then it will have to go out into the wild... maybe then the stranglehold will be broken if not right away then after ... for one thing the projects will have to stop hiding behind their ivory tower walls ... (well, not all of them hide all the time ...)
The saddest thing is that he apparently cannot see what his actions accomplish ... his insistence that teh BOINC Wiki cover only BOINC and nothing about projects or how they interact with BOINC meant I had to stay independent until I could no longer do documentation ... which is silly because BOINC is nothing without the projects and the participant does not really think he deals with BOINC, he deals with projects *USING* BOINC ... a vastly different thing ...
I know developers don't contribute because there is no point ... if he does not want the change they propose, perhaps even code and test (at his suggestion, on the off chance he will accept the change), after awhile they stop coding ... he called for comments on his draft the other day ... I finally took him at his word and sent off a new version that incorporated all of the discussion to that point and he used exactly none of it ...
Those lurking surely see that there is little point in talking, or commenting, or rewriting ... none of it matters, just as you say ...
Onto lighter subjects ... :)
My 9800 GT with 1G VRAM takes up to 17 hours per task as a nominal ... yours likely will be longer ... on the other hand a $400 (and change) GTX 280 takes about 4 hours ... something about spending 4 times as much gets you about 4 times the speed ...
Sadly the server software is still cranked and you have to run dry most of the time before you can refresh your queue which can mean down time if you don't catch it right ... I got woken up from my nap and my 280 was out and as soon as I made the connection I got 5 tasks (1 day buffer).
I will say that, in general, I have been fascinated with the GPU capabilities and find it fun to have something new to tweak and watch. It has given me some brand new things to set goals for ... I was just checking my list and it has 15 entries for this year ... though I just passed one 500K here at Einstein@Home ...
Next up should be WCG over 390K (SaH's number) and Gold badge for CEP (about 4 days short including my pending work (84:06:18), though I have about 4 tasks queued up ... though not enough to get me over the hump yet ... sigh, it looks like it will be mid week before I get there ...
CUDA Tesla Computing
)
Not yet for Einstein@home.
BM
BM
From what I read on the SETI
)
From what I read on the SETI boards Tesla is not yet supported even in SETI. I see a lot of confusion going on CUDA. Also the ATI people would like to use their graphic boards. I think the GPUs are still in beta phase.
Tullio
For a broader discussion on
)
For a broader discussion on GPU support in BOINC projects, see this thread.
CU
Bikeman
At the moment CUDA can only
)
At the moment CUDA can only be used on two projects SaH and GPU Grid ... the SaH implementation has some issues and the general feeling by the mega-crunchers there is that the results returned by the CUDA application have a higher rate of error ...
GPU Grid on the other hand the application is, or appears to be more stable though the current version has a problem with priority and there are server issues on getting work ...
Add in the problems with the work fetch and resource share and you have many potential problems.
If, however, you are not adverse to baby sitting your computer some then going with GPU Grid is not a bad choice at the moment.
On a 9800 GT card the tasks take a minimum of 6 hours with 17 being nominal (rough numbers, it has been a couple days) ... with a GTX 280 it is about 4-6 hours ...
If you are buying a new GPU to process the cost benefit is for the GTX 260 card with the additional processing elements. GPU Grid has some involved discussions ...
In any case, if you want to be an early adopter you can start there and when Einstein gets off the dime you will be set ... :)
Sorry guys, I could not resist ...
Seriously, it is very early days and if you are not up to the hassles of early adoption, then there is no reason not to wait a bit for the dust to settle. THough version 6.5.0 is working reasonably well for me there are serious issues with the BOINC Manager and the way that it does not properly manage resources ...
RE: At the moment CUDA can
)
I believe there is also Folding@Home, although i've never tried it myself.
Bit of an understatement there about "higher rate of error".
Yes the joys of a "development" version. Still it does handle CUDA quite nicely, now if they could fix up the other bits they broke in the process...
BOINC blog
RE: Yes the joys of a
)
NOt sure about the bits they broke in the process, but it does not handle CUDA well ... it lumps the GPU in with the CPUs and you can get a situation where the GPU is dry and BOINC Manager will ask for 0 seconds because the Resource Share and work fetch do not track the GPU tasks separately from the CPU tasks ...
It remains to be seen if any of the other issues raised are addressed at all in the next couple of versions. Based on the trend of the discussion it does not look to be so ...
Anyway, with "fiddling" I have been able to keep my GPUs busy and just went over 100K CS for GPU Grid this evening ... onward to 400,000 ... maybe as soon as the end of the month? Sooner now I have both GPUs working again ...
RE: RE: Yes the joys of a
)
Well 6.5.0 working within the bounds of the current design is better at cuda than 6.4.5. The bits they broke, well don't try to terminate the science apps via BOINC.
I put my 2 cents worth in regarding resource share before attempting to address work-fetch policy. No doubt they will be ignored or put in the "too hard" basket.
BOINC blog
RE: I put my 2 cents worth
)
I think they are being put into the category of: "I am not going to do it because I really don't care" ...
But that is just my bitter, cynical, self speaking ...
None of the issues in BOINC are really all that hard ... there is just a lack of interest in NIH ideas ... so, when we peel back the curtain of time we find all kinds of fixes some of them even coded and tested (I know, I tested them) that never made it into the code base (as far as I know, they may have been snuck in while I was not doing BOINC, though I suspect not because I am still seeing "funky" behavior in task scheduling) ... because Dr. Anderson did not want them.
It mattered not that it addressed problems that were particular with multi-core systems (less evident on dual cores, noticeable on 4 core and better) ... but I digress ...
RE: RE: I put my 2 cents
)
When I was at Seti we physically talked to Dr. A and he was not interested in whatever we had to say. He has a path he is on and will not be deviated from it. IMO it will lead to a different program, other than Boinc, in the future if he does not let change happen. I am not saying this will happen net week, but people will continue to leave if things don't get better at some point.
On a more positive note I bought a 9600 GSO card and it is taking well over 13 hours PER UNIT on Gpugrid. And that is only a projection, it has not actually finished even one unit yet and has run 24 hours! It is in a dual core, 2.66ghz, Vista Home Basic system. I downloaded the latest drivers from Nvidia, but currently it is NOT better!
RE: RE: RE: I put my 2
)
I was pointed to a site by a comment elsewhere and it indicated a BOINC Manager branch that was now being independently maintained. I don't know how good it is or even if it is active at this time ... but I did find that encouraging. I have been coming to the feeling that the only way things are going to look up for BOINC is for it to lose all funding ... then it will have to go out into the wild... maybe then the stranglehold will be broken if not right away then after ... for one thing the projects will have to stop hiding behind their ivory tower walls ... (well, not all of them hide all the time ...)
The saddest thing is that he apparently cannot see what his actions accomplish ... his insistence that teh BOINC Wiki cover only BOINC and nothing about projects or how they interact with BOINC meant I had to stay independent until I could no longer do documentation ... which is silly because BOINC is nothing without the projects and the participant does not really think he deals with BOINC, he deals with projects *USING* BOINC ... a vastly different thing ...
I know developers don't contribute because there is no point ... if he does not want the change they propose, perhaps even code and test (at his suggestion, on the off chance he will accept the change), after awhile they stop coding ... he called for comments on his draft the other day ... I finally took him at his word and sent off a new version that incorporated all of the discussion to that point and he used exactly none of it ...
Those lurking surely see that there is little point in talking, or commenting, or rewriting ... none of it matters, just as you say ...
Onto lighter subjects ... :)
My 9800 GT with 1G VRAM takes up to 17 hours per task as a nominal ... yours likely will be longer ... on the other hand a $400 (and change) GTX 280 takes about 4 hours ... something about spending 4 times as much gets you about 4 times the speed ...
Sadly the server software is still cranked and you have to run dry most of the time before you can refresh your queue which can mean down time if you don't catch it right ... I got woken up from my nap and my 280 was out and as soon as I made the connection I got 5 tasks (1 day buffer).
I will say that, in general, I have been fascinated with the GPU capabilities and find it fun to have something new to tweak and watch. It has given me some brand new things to set goals for ... I was just checking my list and it has 15 entries for this year ... though I just passed one 500K here at Einstein@Home ...
Next up should be WCG over 390K (SaH's number) and Gold badge for CEP (about 4 days short including my pending work (84:06:18), though I have about 4 tasks queued up ... though not enough to get me over the hump yet ... sigh, it looks like it will be mid week before I get there ...